Interesting, I built this system myself while building a toy project (before starting to explore langchain, llamaindex, etc).

What I did was create a short/long summary of each *chunk* of text and then used those to create a long/short summary of the entire document. Then at query time, I feed in either long or short summaries as the context window allows. My next step was to allow the LLM to "double click" into those chunks.